ScaleConf Workshop - General Introduction


These instructions were written for the first Vumi workshop on the 21st of April 2013, right after the ScaleConf conference in Cape Town.

Spotted an error? Please feel free to contribute to the documentation.

What is Vumi?

Vumi is a scalable, multi channel messaging platform. It has been designed to allow large scale mobile messaging in the majority world. It is actively being developed by the Praekelt Foundation and other contributors. It is available as Open Source software under the BSD license.

What were the design goals?

The Praekelt Foundation has a lot of experience building mobile messaging campaigns in the areas such as mobile health, education and democracy. Unfortunately, a lot of this experience comes from having built systems that caused problems in terms of scale and/or maintenance.

Key learnings from these mistakes led to a number of guiding principles in the design of Vumi, such as:

  1. The campaign application logic should be decoupled from how it communicates with the end user.
  2. The campaign application and the means of communication with the end-user should each be re-usable in a different context.
  3. The system should be able to scale by adding more commodity machines, i.e. it should scale horizontally.

The above mentioned guiding principles resulted in a number of core concepts that make up a Vumi application.

A Vumi Message

A Vumi Message is the means of communication inside Vumi. Esentially a Vumi Message is just a bit of JSON that contains information on where a message was received from, who it was addressed to, what the message contents were and some extra metadata to allow it to be routed from and end-user to an application and back again.


Transports provide the communication channel to end users by integrating into various services such as chat systems, mobile network operators or possibly even traditional voice phone lines.

Transports are tasked with translating an inbound request into a standardized Vumi Message and vice-versa.

A simple example would be an SMS, which when received is converted into a bit of JSON that looks something like this:

    "message_id": "message1",
    "to_addr": "1234",
    "from_addr": "27761234567",
    "content": "This is an incoming SMS!",
    "transport_name": "smpp_transport",
    "transport_type": "sms",
    "transport_metadata": {
        // this is a dictionary containing
        // transport specific data


Applications are tasked with either generating messages to be sent to or acting on the messages received from end users via the transports.

As a general rule the Applications should not care about which transport the message was received from, it merely acts on the message contents and provides a suitable reply.

A reply message looks something like this:

    "message_id": "message2",
    "in_reply_to": "message1",
    "to_addr": "27761234567",
    "from_addr": "1234",
    "content": "Thanks! We've received your SMS!",
    "transport_name": "smpp_transport",
    "transport_type": "sms",
    "helper_metadata": {
        // this is a dictionary containing
        // application specific data


Dispatchers are an optional means of connecting Transports and Applications. They allow for more complicated routing between the two.

A simple scenario is an application that receives from a USSD transport but requires the option of also replying via an SMS transport. A dispatcher would allow one to contruct this.

Dispatchers do this by inspecting the messages exchanged between the Transport and the Application and then deciding where it needs to go.

| SMS Transport  |<----+   +------------+    +-------------+
+----------------+     +-->|            |    |             |
                           | Dispatcher |<-->| Application |
+----------------+     +-->|            |    |             |
| USSD Transport |<----+   +------------+    +-------------+

How does it work?

All of these different components are built using the Python programming language using Twisted, an event driven networking library.

The messages between the different components are exchanged and routed using RabbitMQ a high performance AMQP message broker.

For data storage Redis is used for data that are generally temporary but and may potentially be lost. Riak is used for things that need strong availability guarantees.

A sample use case of Redis would be to store session state whereas Riak would be used to store all messages sent and received indefinitely.

Supervisord is used to manage all the different processes and provide any easy commandline tool to start and stop them.

Let’s get started!

As part of the workshop we will provide you with a South African USSD code and an SMS longcode. In the next section we’ll help you get Vumi running on your local machine so you can start developing your first application!